Quoting: TheRiddickHowever gamedevs still need to send off textures to NVIDIA for processing, but maybe one day developers will have access to those tools themselves and NVIDIA won't need to be involved (we can only hope).

They don't.

The AI for DLSS 1 needed to be trained on "what this game should look like" to provide its results, and the per-game magic numbers were included along with the per-game optimisations in the fat Windows driver.

The AI for DLSS 2 is trained on "what games look like" and the one set of magic numbers is used for everything.

Devs need to include motion vectors and the previous upscaled frame in their engine to use the DLSS mechanism for the next frame.

View this comment - View article - View full comments

Quoting: x_wingIIRC, the first sample of Physx I saw was on 2005 and it was from the former company that created the tech, using dedicated hardware, which was in a very early stage (I'm almost sure that their dedicated solution never got to the market).

The PPUs definitely existed. I doubt that many got sold, because the business case for them was rubbish, but you could get pre-built gaming machines with them in. The technology was also in a bunch of console games before Nvidia bought Ageia.

In the moment that Nvidia bought that company, their strategy was to implement that solution into the GPU. So, Nvidia wanted to move physics calculation into GPU as use case of GPGPU.

Of course they did. Buying an extra PPU was silly, but GPGPU is great. And of course they wanted it to be a market differentiator to make back the purchase price, particularly as Intel had just bought Havok at the time.

View this comment - View article - View full comments

Quoting: x_wingSo, the idea was to accelerate physics execution using the GPU but their reluctance to make a standard made them fail and 15 years after they first release of Physx we are still using the CPU.

No, the idea was that you'd buy a separate card just for accelerating physics calculations. But that was silly: no one was going to buy a card just for that, and no one was going to put support into their game for something that no one had. So Nvidia bought the company and made it so that you could run those calculations on the GPU that you already had. Then they open sourced it some time later.

View this comment - View article - View full comments

Quoting: GuestAs far as I know, FSR is basically a couple of shaders? So I wonder if it could be injected to any game via something like vkBasalt. Pretty sure it wouldn't be quite something so simple. Suppose I should read up on it more to know if that would be viable. It'd definitely be very cool if it was, just from a technical perspective, even if I wouldn't really benefit much personally from this kind of thing.

If it has some particularly tasty rescaling algorithms, a nicer place to pick it up would be gamescope if, say, there were some performance-constrained AMD hardware that was habitually upscaling.

View this comment - View article - View full comments

Added gsp.bin firmware file which is used to offload the GPU initialization and management tasks on some GPUs.

I wonder if this will help or hurt reclocking with nouveau. Hopefully the former.

View this comment - View article - View full comments

Quoting: slaapliedjeGranted, nvidia doesn't support Mac at all, which I still find amusing.

They can't. On Windows and Linux, the GPU vendor provides the API implementation. On Macs, Apple do. Apple and Nvidia had a falling out, so no more Nvidia hardware in Macs, so no support from Apple for Nvidia hardware in Macs.

View this comment - View article - View full comments

One thing that would be interesting is if Intel licensed the DLSS tech from Nvidia. They do have the hardware, since they're chasing the same machine learning market as Nvidia. Tensor cores on consumer products is essentially a lemons -> lemonade situation. On the one hand they won't want to give Intel a leg up in the dedicated GPU market, nor in the very lucrative machine learning market, but on the other hand a feature that both Nvidia and Intel have that AMD doesn't will put the squeeze on their mutual competitor.

View this comment - View article - View full comments

Quoting: ShmerlSo instead of being essentially a fancy display, the headset will have its own GPU?

That moves things into interesting direction of eventually rendering everything on the headset itself, but I guess they need to get small enough first while being powerful. Current GPUs required for VR use a ton of power and need serious heat dissipation.

Standalone VR headsets already exist with their own GPU. The Oculus Quest uses the Adreno 540 and the Quest 2 uses the Adreno 650. In terms of graphics power they're pretty weak, but going stronger risks melting your face. The idea of this is that you can offload some of the face melting stuff to a different machine, but have enough local processing power that the latency doesn't make you hurl.

View this comment - View article - View full comments

Quoting: gradyvuckovic

Quoting: CatKillerSo, I think the idea is to use the same kind of tech as game streaming; the computer renders the scene based on position information from the headset, and sends essentially a 360° video stream to the headset, which can be freely navigated using the headset. If there's enough bandwidth, that could be quite interesting.

Ya know, if they included depth information per pixel, it would be possible to not only correctly correct the 360 image to adjust for head rotation, but also give proper depth displacement to the 360 video to account for head movement with near zero latency as well. Would hide a lot of latency or frame rate issues in most circumstances.

So, looking at the patent, that seems to be something that's also included in the stream. The headset sends movement data, and the computer guesstimates where the headset will be at render time based on that. The stream it sends to the headset includes pixel data (so, the video stream) as well as pose data, depth data, motion vector data, parallax occlusion data, and "extra pixel data." So if the next frame is late, or the guesstimate is a bit wrong, the headset has enough data to correct for it. Could be neat.

Also, the computer seems to render the scene, but the headset renders the player model, which should also reduce any latency discrepancy between what the player's doing and what they're seeing.

View this comment - View article - View full comments

So, I think the idea is to use the same kind of tech as game streaming; the computer renders the scene based on position information from the headset, and sends essentially a 360° video stream to the headset, which can be freely navigated using the headset. If there's enough bandwidth, that could be quite interesting.

View this comment - View article - View full comments